Leveraging the Latent Diffusion Models for Offline Facial Multiple Appropriate Reactions Generation